We now have a stream of packets that are stored in files that originated as frames about glove state. There are approximately 25 such frames per second of signing. While it might be possible for this information to be fed directly into some learning algorithm (in fact, it was attempted by Murakami and Taguchi [MT91]), this is likely to be computationally expensive and does not take advantage of good-discriminating attributes that can be extracted simply from the data. Furthermore, a great deal of effort would need to be expended in extending existing learning algorithms to fulfil this obligations -- since most common learning algorithms are not able to handle time-varying data in a sensible manner.
Thus in this investigation, the focus was on extracting pieces of
information from the data, which we will term features or
attributes
that are good for discriminating between
signs.
This, of course, is the central part of the investigation, and most of the effort was expended in trying to come up with, calculate, assess and validate features that can be extracted by simple processing.